On 18/6/26 01:41, Jason Gunthorpe wrote:

On Wed, Jun 17, 2026 at 10:50:39AM +1000, Alexey Kardashevskiy wrote:

quoted

quoted

@@ -193,16 +193,31 @@ void *dma_direct_alloc(struct device *dev, size_t size,
   		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
   {
   	bool remap = false, set_uncached = false;
-	bool mark_mem_decrypt = true;
+	bool mark_mem_decrypt = false;
   	struct page *page;
   	void *ret;
+	/*
+	 * DMA_ATTR_CC_SHARED is not a caller-visible dma_alloc_*()
+	 * attribute. The direct allocator uses it internally after it has
+	 * decided that the backing pages must be shared/decrypted, so the
+	 * rest of the allocation path can consistently select DMA addresses,
+	 * choose compatible pools and restore encryption on free.

Why this limit?

Context: I am looking for a memory pool for a few shared pages (to
do some guest<->host communication), SWIOTLB seems like the right
fit but swiotlb_alloc() is not exported and
dma_direct_alloc(DMA_ATTR_CC_SHARED) is not allowed.  Thanks,

Then setup your struct device so that the DMA API knows the
guest<->host channel requires unecrypted and it will work correctly.

I think this is a reasonable API to use for that, and I was just
advocating that hyperv should be using it too.

But it all relies on a properly setup struct device.

Sounds good but how do I do that in practice? DMA_ATTR_CC_SHARED is not externally available so I'll have to trick the DMA layer into using SWIOTLB (which is still all shared, right?) as I specifically want to skip page conversions. Setting low DMA mask won't guarantee that the DMA layer won't allocate a page outside of SWIOTLB and convert it. Manually do

dev->dma_io_tlb_mem->force_bounce = true;
dev->dma_io_tlb_mem->for_allow = true;

?
Or follow the Aneesh'es genpool approach? Thanks,

Jason

-- 
Alexey